2 * Copyright (C) 2016 Facebook
3 * Copyright (C) 2013-2014 Jens Axboe
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program. If not, see <https://www.gnu.org/licenses/>.
18 #include <linux/sched.h>
19 #include <linux/random.h>
20 #include <linux/sbitmap.h>
21 #include <linux/seq_file.h>
23 int sbitmap_init_node(struct sbitmap
*sb
, unsigned int depth
, int shift
,
24 gfp_t flags
, int node
)
26 unsigned int bits_per_word
;
30 shift
= ilog2(BITS_PER_LONG
);
32 * If the bitmap is small, shrink the number of bits per word so
33 * we spread over a few cachelines, at least. If less than 4
34 * bits, just forget about it, it's not going to work optimally
38 while ((4U << shift
) > depth
)
42 bits_per_word
= 1U << shift
;
43 if (bits_per_word
> BITS_PER_LONG
)
48 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
55 sb
->map
= kzalloc_node(sb
->map_nr
* sizeof(*sb
->map
), flags
, node
);
59 for (i
= 0; i
< sb
->map_nr
; i
++) {
60 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
61 depth
-= sb
->map
[i
].depth
;
65 EXPORT_SYMBOL_GPL(sbitmap_init_node
);
67 void sbitmap_resize(struct sbitmap
*sb
, unsigned int depth
)
69 unsigned int bits_per_word
= 1U << sb
->shift
;
73 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
75 for (i
= 0; i
< sb
->map_nr
; i
++) {
76 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
77 depth
-= sb
->map
[i
].depth
;
80 EXPORT_SYMBOL_GPL(sbitmap_resize
);
82 static int __sbitmap_get_word(unsigned long *word
, unsigned long depth
,
83 unsigned int hint
, bool wrap
)
85 unsigned int orig_hint
= hint
;
89 nr
= find_next_zero_bit(word
, depth
, hint
);
90 if (unlikely(nr
>= depth
)) {
92 * We started with an offset, and we didn't reset the
93 * offset to 0 in a failure case, so start from 0 to
96 if (orig_hint
&& hint
&& wrap
) {
103 if (!test_and_set_bit(nr
, word
))
107 if (hint
>= depth
- 1)
114 int sbitmap_get(struct sbitmap
*sb
, unsigned int alloc_hint
, bool round_robin
)
116 unsigned int i
, index
;
119 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
121 for (i
= 0; i
< sb
->map_nr
; i
++) {
122 nr
= __sbitmap_get_word(&sb
->map
[index
].word
,
123 sb
->map
[index
].depth
,
124 SB_NR_TO_BIT(sb
, alloc_hint
),
127 nr
+= index
<< sb
->shift
;
131 /* Jump to next index. */
133 alloc_hint
= index
<< sb
->shift
;
135 if (index
>= sb
->map_nr
) {
143 EXPORT_SYMBOL_GPL(sbitmap_get
);
145 int sbitmap_get_shallow(struct sbitmap
*sb
, unsigned int alloc_hint
,
146 unsigned long shallow_depth
)
148 unsigned int i
, index
;
151 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
153 for (i
= 0; i
< sb
->map_nr
; i
++) {
154 nr
= __sbitmap_get_word(&sb
->map
[index
].word
,
155 min(sb
->map
[index
].depth
, shallow_depth
),
156 SB_NR_TO_BIT(sb
, alloc_hint
), true);
158 nr
+= index
<< sb
->shift
;
162 /* Jump to next index. */
164 alloc_hint
= index
<< sb
->shift
;
166 if (index
>= sb
->map_nr
) {
174 EXPORT_SYMBOL_GPL(sbitmap_get_shallow
);
176 bool sbitmap_any_bit_set(const struct sbitmap
*sb
)
180 for (i
= 0; i
< sb
->map_nr
; i
++) {
186 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set
);
188 bool sbitmap_any_bit_clear(const struct sbitmap
*sb
)
192 for (i
= 0; i
< sb
->map_nr
; i
++) {
193 const struct sbitmap_word
*word
= &sb
->map
[i
];
196 ret
= find_first_zero_bit(&word
->word
, word
->depth
);
197 if (ret
< word
->depth
)
202 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear
);
204 unsigned int sbitmap_weight(const struct sbitmap
*sb
)
206 unsigned int i
, weight
= 0;
208 for (i
= 0; i
< sb
->map_nr
; i
++) {
209 const struct sbitmap_word
*word
= &sb
->map
[i
];
211 weight
+= bitmap_weight(&word
->word
, word
->depth
);
215 EXPORT_SYMBOL_GPL(sbitmap_weight
);
217 void sbitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
219 seq_printf(m
, "depth=%u\n", sb
->depth
);
220 seq_printf(m
, "busy=%u\n", sbitmap_weight(sb
));
221 seq_printf(m
, "bits_per_word=%u\n", 1U << sb
->shift
);
222 seq_printf(m
, "map_nr=%u\n", sb
->map_nr
);
224 EXPORT_SYMBOL_GPL(sbitmap_show
);
226 static inline void emit_byte(struct seq_file
*m
, unsigned int offset
, u8 byte
)
228 if ((offset
& 0xf) == 0) {
231 seq_printf(m
, "%08x:", offset
);
233 if ((offset
& 0x1) == 0)
235 seq_printf(m
, "%02x", byte
);
238 void sbitmap_bitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
241 unsigned int byte_bits
= 0;
242 unsigned int offset
= 0;
245 for (i
= 0; i
< sb
->map_nr
; i
++) {
246 unsigned long word
= READ_ONCE(sb
->map
[i
].word
);
247 unsigned int word_bits
= READ_ONCE(sb
->map
[i
].depth
);
249 while (word_bits
> 0) {
250 unsigned int bits
= min(8 - byte_bits
, word_bits
);
252 byte
|= (word
& (BIT(bits
) - 1)) << byte_bits
;
254 if (byte_bits
== 8) {
255 emit_byte(m
, offset
, byte
);
265 emit_byte(m
, offset
, byte
);
271 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show
);
273 static unsigned int sbq_calc_wake_batch(unsigned int depth
)
275 unsigned int wake_batch
;
278 * For each batch, we wake up one queue. We need to make sure that our
279 * batch size is small enough that the full depth of the bitmap is
280 * enough to wake up all of the queues.
282 wake_batch
= SBQ_WAKE_BATCH
;
283 if (wake_batch
> depth
/ SBQ_WAIT_QUEUES
)
284 wake_batch
= max(1U, depth
/ SBQ_WAIT_QUEUES
);
289 int sbitmap_queue_init_node(struct sbitmap_queue
*sbq
, unsigned int depth
,
290 int shift
, bool round_robin
, gfp_t flags
, int node
)
295 ret
= sbitmap_init_node(&sbq
->sb
, depth
, shift
, flags
, node
);
299 sbq
->alloc_hint
= alloc_percpu_gfp(unsigned int, flags
);
300 if (!sbq
->alloc_hint
) {
301 sbitmap_free(&sbq
->sb
);
305 if (depth
&& !round_robin
) {
306 for_each_possible_cpu(i
)
307 *per_cpu_ptr(sbq
->alloc_hint
, i
) = prandom_u32() % depth
;
310 sbq
->wake_batch
= sbq_calc_wake_batch(depth
);
311 atomic_set(&sbq
->wake_index
, 0);
313 sbq
->ws
= kzalloc_node(SBQ_WAIT_QUEUES
* sizeof(*sbq
->ws
), flags
, node
);
315 free_percpu(sbq
->alloc_hint
);
316 sbitmap_free(&sbq
->sb
);
320 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
321 init_waitqueue_head(&sbq
->ws
[i
].wait
);
322 atomic_set(&sbq
->ws
[i
].wait_cnt
, sbq
->wake_batch
);
325 sbq
->round_robin
= round_robin
;
328 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node
);
330 void sbitmap_queue_resize(struct sbitmap_queue
*sbq
, unsigned int depth
)
332 unsigned int wake_batch
= sbq_calc_wake_batch(depth
);
335 if (sbq
->wake_batch
!= wake_batch
) {
336 WRITE_ONCE(sbq
->wake_batch
, wake_batch
);
338 * Pairs with the memory barrier in sbq_wake_up() to ensure that
339 * the batch size is updated before the wait counts.
341 smp_mb__before_atomic();
342 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++)
343 atomic_set(&sbq
->ws
[i
].wait_cnt
, 1);
345 sbitmap_resize(&sbq
->sb
, depth
);
347 EXPORT_SYMBOL_GPL(sbitmap_queue_resize
);
349 int __sbitmap_queue_get(struct sbitmap_queue
*sbq
)
351 unsigned int hint
, depth
;
354 hint
= this_cpu_read(*sbq
->alloc_hint
);
355 depth
= READ_ONCE(sbq
->sb
.depth
);
356 if (unlikely(hint
>= depth
)) {
357 hint
= depth
? prandom_u32() % depth
: 0;
358 this_cpu_write(*sbq
->alloc_hint
, hint
);
360 nr
= sbitmap_get(&sbq
->sb
, hint
, sbq
->round_robin
);
363 /* If the map is full, a hint won't do us much good. */
364 this_cpu_write(*sbq
->alloc_hint
, 0);
365 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
366 /* Only update the hint if we used it. */
368 if (hint
>= depth
- 1)
370 this_cpu_write(*sbq
->alloc_hint
, hint
);
375 EXPORT_SYMBOL_GPL(__sbitmap_queue_get
);
377 int __sbitmap_queue_get_shallow(struct sbitmap_queue
*sbq
,
378 unsigned int shallow_depth
)
380 unsigned int hint
, depth
;
383 hint
= this_cpu_read(*sbq
->alloc_hint
);
384 depth
= READ_ONCE(sbq
->sb
.depth
);
385 if (unlikely(hint
>= depth
)) {
386 hint
= depth
? prandom_u32() % depth
: 0;
387 this_cpu_write(*sbq
->alloc_hint
, hint
);
389 nr
= sbitmap_get_shallow(&sbq
->sb
, hint
, shallow_depth
);
392 /* If the map is full, a hint won't do us much good. */
393 this_cpu_write(*sbq
->alloc_hint
, 0);
394 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
395 /* Only update the hint if we used it. */
397 if (hint
>= depth
- 1)
399 this_cpu_write(*sbq
->alloc_hint
, hint
);
404 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow
);
406 static struct sbq_wait_state
*sbq_wake_ptr(struct sbitmap_queue
*sbq
)
410 wake_index
= atomic_read(&sbq
->wake_index
);
411 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
412 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
414 if (waitqueue_active(&ws
->wait
)) {
415 int o
= atomic_read(&sbq
->wake_index
);
418 atomic_cmpxchg(&sbq
->wake_index
, o
, wake_index
);
422 wake_index
= sbq_index_inc(wake_index
);
428 static void sbq_wake_up(struct sbitmap_queue
*sbq
)
430 struct sbq_wait_state
*ws
;
431 unsigned int wake_batch
;
435 * Pairs with the memory barrier in set_current_state() to ensure the
436 * proper ordering of clear_bit()/waitqueue_active() in the waker and
437 * test_and_set_bit()/prepare_to_wait()/finish_wait() in the waiter. See
438 * the comment on waitqueue_active(). This is __after_atomic because we
439 * just did clear_bit() in the caller.
441 smp_mb__after_atomic();
443 ws
= sbq_wake_ptr(sbq
);
447 wait_cnt
= atomic_dec_return(&ws
->wait_cnt
);
449 wake_batch
= READ_ONCE(sbq
->wake_batch
);
451 * Pairs with the memory barrier in sbitmap_queue_resize() to
452 * ensure that we see the batch size update before the wait
455 smp_mb__before_atomic();
457 * If there are concurrent callers to sbq_wake_up(), the last
458 * one to decrement the wait count below zero will bump it back
459 * up. If there is a concurrent resize, the count reset will
460 * either cause the cmpxchg to fail or overwrite after the
463 atomic_cmpxchg(&ws
->wait_cnt
, wait_cnt
, wait_cnt
+ wake_batch
);
464 sbq_index_atomic_inc(&sbq
->wake_index
);
469 void sbitmap_queue_clear(struct sbitmap_queue
*sbq
, unsigned int nr
,
472 sbitmap_clear_bit(&sbq
->sb
, nr
);
474 if (likely(!sbq
->round_robin
&& nr
< sbq
->sb
.depth
))
475 *per_cpu_ptr(sbq
->alloc_hint
, cpu
) = nr
;
477 EXPORT_SYMBOL_GPL(sbitmap_queue_clear
);
479 void sbitmap_queue_wake_all(struct sbitmap_queue
*sbq
)
484 * Pairs with the memory barrier in set_current_state() like in
488 wake_index
= atomic_read(&sbq
->wake_index
);
489 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
490 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
492 if (waitqueue_active(&ws
->wait
))
495 wake_index
= sbq_index_inc(wake_index
);
498 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all
);
500 void sbitmap_queue_show(struct sbitmap_queue
*sbq
, struct seq_file
*m
)
505 sbitmap_show(&sbq
->sb
, m
);
507 seq_puts(m
, "alloc_hint={");
509 for_each_possible_cpu(i
) {
513 seq_printf(m
, "%u", *per_cpu_ptr(sbq
->alloc_hint
, i
));
517 seq_printf(m
, "wake_batch=%u\n", sbq
->wake_batch
);
518 seq_printf(m
, "wake_index=%d\n", atomic_read(&sbq
->wake_index
));
520 seq_puts(m
, "ws={\n");
521 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
522 struct sbq_wait_state
*ws
= &sbq
->ws
[i
];
524 seq_printf(m
, "\t{.wait_cnt=%d, .wait=%s},\n",
525 atomic_read(&ws
->wait_cnt
),
526 waitqueue_active(&ws
->wait
) ? "active" : "inactive");
530 seq_printf(m
, "round_robin=%d\n", sbq
->round_robin
);
532 EXPORT_SYMBOL_GPL(sbitmap_queue_show
);